Carebears hasn't been covered enough yet. Precision strikes have taken place but addition bear bombing can be done.

0.2

Fixed and broke it all at the same time. Learned how far you can push a small dataset, this is how far. You can see the generation errors from squeezing every image as hard as possible. Thankfully I just found a way better version of the care bears I needed to create most of em, this is the last version on this dataset but the new set will use basically exactly the settings of this one so it's a good point to post and it's WAY more stable when feeding it into checkpoints and crossing the streams with.

OH and despite having separate training folders now, it gets confused with color so you have to force it and even then depending on checkpoints it likes sticking. The newer dataset will have a better set including the tummy symbols, which are stupid hit and miss, mostly miss.

I want Shreeky next.

0.1 testing text and image coverage.